Long-form text is where detection actually works
Most pages on this site spend their first paragraph telling you what detection can't do. This one gets the good case: articles, blog posts, and newsletters are hundreds to thousands of words long, and length is what statistical detection feeds on. At 500+ words, small regularities — how predictable each word is, how uniform the rhythm stays across sections — accumulate into real evidence. This is our strongest measured territory: paste the piece into the free checker and you get a calibrated probability whose error rate at that length is published, with raw counts, not asserted.
The “polished but soulless” feeling, made honest
Readers describe the same tells: prose that is fluent and weirdly frictionless, sections that summarize themselves, hedged claims with no author anywhere in them — and the famous vocabulary (“delve,” “tapestry,” “in today's fast-paced world”). Here is the honest frame for those lists: any single word is worthless as evidence, because humans use them all. What makes word choice meaningful is aggregation across a whole text — hundreds of small word-level regularities weighed together. That is not a metaphor for what our scan does; it is literally what the scan does, with calibration so the output probability means what it says. Your instinct reads the same surface — the scan just counts it honestly, and tells you how often that count is wrong.
Comments and posts: the floor still applies
The moment you leave long-form, the length floor returns. A 15-word comment cannot be classified by anyone — including us — so for social feeds, shift to account-level signals: posting cadence no human keeps, engagement that doesn't match follower counts, and interchangeable phrasing across many comments. That last one is checkable: paste a pile of comments from the same account as one batch, exactly like the review batch method, once they total roughly 150 words. Profile pictures have their own deterministic check — read the file's Content Credentials rather than squinting at the earrings.
The dead-internet question, with real numbers
“Is everything AI now?” deserves a calibrated answer too. What is actually measured: industry bot-traffic reports (Imperva's annual Bad Bot Report among them) have put automated traffic near half of all web requests — but that counts crawlers, scrapers, and monitoring, not AI-written articles. How much visible content is AI-generated has no trustworthy measurement; the viral percentages are extrapolations from small samples. The honest position is between the extremes: AI text is now a large and growing share of new publishing, the feeling that “every article sounds the same” has a real statistical basis — and the internet still contains humans, writing. Check the piece in front of you instead of trusting a vibe about the whole web.
What a verdict does — and does not — tell you
A likely-AI result estimates statistical origin. It does not tell you the piece is wrong, low-effort, or deceptive: plenty of accurate, well-edited journalism now starts from an AI draft. A likely-human result doesn't certify truth either — humans wrote every hoax in history. Detection answers one question well; fact-checking, bylines that stand behind claims, and sources you can follow answer the one you usually care about. Why detectors disagree with each other — and what an honest accuracy claim looks like — is covered in Can AI detectors be trusted?
What Cobalynx can and can't do here
We detect statistical origin, not truth or quality: an AI-drafted article can be accurate and human-edited, and a fully human one can be wrong. Long-form text gets our best-measured accuracy — the bands are on the evidence page — while short comments stay below any detector's floor, and we say “too short” rather than guess. AI-assisted writing is now normal; the interesting question is whether anyone checked the facts.
Sources
- Imperva, annual Bad Bot Report — automated share of web traffic (requests, not content; accessed Aug 2026).
- Weber-Wulff et al., “Testing of detection tools for AI-generated text,” International Journal for Educational Integrity (2023), arxiv.org/abs/2306.15666 — detector performance varies with text length and editing.
Common questions
Are “ChatGPT words” like delve and tapestry proof of AI?
No single word is — humans use every word on those viral lists. Word choice becomes evidence only in statistical aggregate across a whole text, which is precisely what a calibrated scan measures and a vibe check cannot (how the scoring works). Treat any one-word gotcha as entertainment, not analysis.
Can I check a single social media comment?
Not reliably — a 15-word comment is far below any detector's floor, ours included. Account-level signals work better: posting cadence, engagement ratios, and sameness across many comments from the same account, which you can paste together as one batch once they total roughly 150 words. Profile pictures can be checked for generator credentials.
How much of the internet is AI-generated?
Nobody has a trustworthy measurement of AI-written content, and figures you see quoted are extrapolations. What is measured: industry bot-traffic reports (Imperva's annual Bad Bot Report among them) have put automated traffic near half of all web requests — but that counts crawlers and scrapers, not AI-written articles. Be suspicious of any precise-sounding percentage, including in AI-generated answers to this question.
Does an AI verdict mean the article is untrustworthy?
No — the verdict estimates statistical origin, not truth or quality. An AI-drafted article can be accurate, human-edited, and useful; a fully human one can be wrong or dishonest. AI assistance is now normal in publishing, so the sharper questions are whether anyone checked the facts and whether the byline stands behind them.